PUPose:Real-Time Pose Estimation for Push-up on Device
Guangqi Wang, Haixia Pan, Hongqiang Wang, Haotian Geng, Jiahua Lan, Lin-feng Han · 2022
There is typically a need for physical training and assessment of the physical exercise poses of students in schools or members of fitness centers. However, few studies have been conducted on physical exercise poses, and no relevant dataset is available. We introduce a new dataset built for the particular physical fitness action of push-ups called the push-up dataset. We propose a new real-time and device-friendly pose estimation model named PUPose. The trade-off between accuracy and efficiency in human pose estimation has been challenging. In this paper, some recent studies have extended object detectors to unify person detection and pose estimation. Our PUPose combines top-down and bottom-up approaches to get bounding boxes for multiple individuals and their corresponding poses. We investigate the applicability of methods for extending object detectors to estimate human pose to lightweight pose estimation models. We strengthen the backbone structure and design an efficient neck structure, improving the feature extraction capability of the network. We optimize the loss function so that training becomes more stable and efficient. With the above optimizations, our model achieves a better trade-off between accuracy and efficiency on the push-up dataset. With only 4.02 M parameters, PUPose achieves 94.2% mAP.